Which experiment would change the next decision? An evaluation brief for public-sector services.
Which experiment would change the next decision?
A good evaluation starts with a consequential uncertainty. The goal is to produce evidence that informs an actual choice, not simply to demonstrate that a feature can run. A public-sector team is evaluating an assistant that organizes guidance and prepares draft answers for staff review. Published guidance can change, and a confident answer can obscure the source, applicability conditions, or authority behind a statement.
A practical starting point
For this evaluation, name the decision before writing the test and identify the result that would cause the team to stop. State the hypothesis, success criteria, baseline, and operating conditions before the test. Choose representative material and include a plausible failure case. Identify the resources and dependencies that influence whether the result could transfer to a broader setting. The immediate concern is whether an answer derived from superseded guidance is presented as the current basis for a consequential decision. The review should make that possibility testable, rather than relying on the apparent fluency or completeness of the output.
Examine the boundary
Include a retired guidance document alongside its replacement and ask a question whose answer depends on the revision. Examine the result against the original question. A successful demonstration may depend on careful input selection, hidden manual work, or conditions that will not hold in operation. Document those dependencies and decide what needs a further test. Bring the service owner and the authority responsible for the underlying guidance into the review when the finding affects an operational or institutional decision. Their role is to connect the evidence with the authority needed to act on it.
Keep the evidence connected
Use a response record showing the applicable source version, cited passage, unresolved question, and reviewing staff role. Record the hypothesis, baseline, test material, configuration, observed behavior, deviations, and the decision the result supports. A later reviewer should be able to see the original question, the observations that mattered, and the point at which the team moved from investigation to a decision. Preserve contradictions and unresolved questions alongside the outcome.
From evaluation to use
Use the assistant to support preparation and review. Establish how staff identify authoritative guidance and handle cases that cannot be resolved from the available record. A bounded evaluation supports a bounded conclusion. It does not establish performance across every user, environment, data source, or future system version. The practical next step is a bounded review with an identified owner, a stated question, and an evidence package that supports the decision.
Continue the conversation
Bring your operating question.
Connect your objective with the relevant attribution, governance, research, or licensing pathway.
Contact Spyris
